Kimi K2 is the open model that put Moonshot AI on the map in July 2025. It was built for coding and for tasks where the AI works by itself with tools. Moonshot has since retired it from its own service, but anyone can still download and run it. Here is what it was, where it stands, and what to use today.
Kimi K2 had a short, busy life: three main versions in under a year.
A very large open model for coding and agent tasks, released in July 2025.
Retired from Moonshot's API. The weights are still free to download.
Kimi K3 for the hardest work, Kimi K2.6 for everyday jobs at a low price.
A model is the "brain" behind an AI assistant. Kimi K2 was Moonshot's first widely noticed open one. "Open" means Moonshot published the model itself, so other people could download it, study it and run it on their own machines.
Kimi K2 was built for doing, more than chatting. Moonshot described it as tuned for coding and for "agentic" work, where the model calls tools, reads the results and keeps going until a task is finished. The first version answered straight away, without long reasoning.
Figures from Moonshot's Kimi K2 model cards on Hugging Face. New to these words? See how Kimi works.
| Version | When | Context | What was new |
|---|---|---|---|
| Kimi K2 (0711) | July 2025 | 128K | The original. Fast answers, strong at coding and tool use |
| Kimi K2 (0905) | September 2025 | 256K | Twice the context, better coding, better-looking front ends |
| Kimi K2 Thinking | Late 2025 | 256K | Reasons step by step. Moonshot says it stays stable across 200 to 300 tool calls in a row |
On a small screen, swipe the table sideways to see every column.
The best-known test for coding models is SWE-bench Verified. It asks a model to fix real bugs from real software projects. The score is the share it fixes.
These are Moonshot's own published results from its model cards, mostly run with its own test setup. We did not run them. They are useful for seeing the trend between versions, less so for comparing with other companies' models.
Each version fixed more bugs than the last. Newer Kimi models have since moved well past these figures. See the Kimi comparison hub.
| Where | Status | What it means |
|---|---|---|
| Kimi app and kimi.com | No | The app runs on current models such as Kimi K3 and Kimi K2.6 |
| Moonshot's API | Retired | Discontinued on 25 May 2026. All five kimi-k2 model names are listed as deprecated |
| Your own servers | Yes | The weights are on Hugging Face. You need powerful server hardware and the skills to run it |
| Other hosting companies | Varies | Some still offer it and some have removed it. Check with the provider |
On a small screen, swipe the table sideways to see every column.
Moonshot's documentation recommends moving to Kimi K3. Depending on the job and your budget, one of the other current models may fit better.
The top model and Moonshot's suggested replacement. Reads about 1 million tokens, and understands images and video.
API: $3 in / $15 out per million tokens · Kimi K3 guide
The current coding model. The closest match if you used Kimi K2 mainly for programming.
API: $0.95 in / $4 out · K2.7 Code guide
The everyday general model at a low price. Thinking can be switched on or off.
API: $0.95 in / $4 out · Kimi K2.6 guide
The K3 recommendation is Moonshot's. The other two suggestions are ours, based on what each model is built for. Prices from Moonshot's price page, before tax.
If your code still calls a Kimi K2 model on Moonshot's API, it needs a new model name. These are the retired names.
| Retired name | Consider instead |
|---|---|
kimi-k2-0711-preview | kimi-k2.6 or kimi-k3 |
kimi-k2-0905-preview | kimi-k2.7-code for coding, or kimi-k2.6 |
kimi-k2-turbo-preview | kimi-k2.7-code-highspeed |
kimi-k2-thinking | kimi-k3 |
kimi-k2-thinking-turbo | kimi-k3 |
Retired names from Moonshot's models page. The pairings are our suggestions. Moonshot's own advice is simply to use the latest model, kimi-k3.
Help: Kimi API setup · Kimi K3 API · Kimi K2.6 API · API pricing
| Kimi K2 | Kimi K2.6 | Kimi K3 | |
|---|---|---|---|
| Released | July 2025 | 2026 | July 2026 |
| Reads at once | 128K to 256K tokens | About 262K tokens | About 1 million tokens |
| Thinking | Only in the Thinking version | On or off | Always on |
| Size | 1 trillion parameters | Not covered here | 2.8 trillion parameters |
| On Moonshot's API | Retired | Yes | Yes |
| Best for today | Research and self-hosting | Everyday work at low cost | The hardest tasks |
On a small screen, swipe the table sideways to see every column.
More comparisons: Kimi K2 vs K2.5 · Kimi K2.5 · choose a Kimi model